Machine Learning for Big Data

Deployment in Organizations

Session 9 | 2026-10-02

• Explain the strategic role of data analytics in data-driven organizations.
• Discuss the ethical and legal boundaries of data analytics in organizations.

Industry-level perspectives

Does the deployment of analytics pay off?

Müller et al. (2018): The Effect of Big Data and Analytics on Firm Performance

The study examines whether firms with live big data and analytics (BDA) assets show higher productivity than comparable firms without such assets.

Empirical setting

• 814 publicly traded firms
• 2008–2014 panel data
• 5,698 firm-year observations
• Objective data on live BDA assets
• Financial data from Compustat

Outcome and controls

• Outcome: \(\log(Sales)\)
• Inputs: labor, capital, materials
• Controls: General IT, industry, and year
• Moderators: IT intensity and competition

\[ \begin{aligned} \log(Sales) =\;& \beta_0 + \beta_1 \log(Labor) + \beta_2 \log(Capital) + \beta_3 \log(Materials) \\ &+ \beta_4 BDA + Controls + \varepsilon \end{aligned} \]

Cobb–Douglas-style production function log-linear model (not a logistic regression) Fixed effects: each firm has its own intercept (addressing omitted variable bias)

The study follows the IT business value tradition and uses a Cobb–Douglas production function. BDA is operationalized as ownership of live BDA assets, not as a self-reported perception of analytics maturity.

It is challenging to isolate the effects of analytics from those of AI or IT more broadly.

Findings

	OLS	FE
log(Capital)	0.090*** (0.015)	0.127*** (0.029)
log(Labor)	0.296*** (0.024)	0.472*** (0.043)
log(Materials)	0.667*** (0.023)	0.442*** (0.035)
BDA	0.041** (0.020)	0.038* (0.021)
Industry dummies?	Yes	Yes
Year dummies?	Yes	Yes
IT Asset dummies?	Yes	Yes
Observations	5,698	5,698
\(R^2\)	0.977	0.791
Adjusted \(R^2\)	0.977	0.755

Key findings and implications

• Analytics assets alone are insufficient: value depends on complementary resources such as data infrastructure, enterprise systems, and analytical skills.
• Industry conditions matter: IT-intensive and highly competitive industries show clearer productivity effects.
• Causality remains difficult: observational firm-level data can reduce, but not fully eliminate, concerns about reverse causality and omitted variables.

At the industry level, analytics deployment should be understood as part of a broader configuration of digital capabilities, competitive pressure, and organizational capacity to turn data into decisions.

The high \(R^2\) is expected because \(\log(Sales)\) is modeled with major production inputs, especially \(\log(Materials)\), \(\log(Labor)\), and \(\log(Capital)\), plus industry/year controls. It should not be interpreted as “BDA explains 97.7% of sales.”

A meta-analysis

Oesterreich et al. (2022): What translates big data into business value?

This meta-analysis examines how differences in business analytics (BA) resources, capabilities, and contextual factors are associated with differences in firm performance.

Empirical setting

• 125 firm-level studies
• 123 articles
• 2012–2021 research period
• 26 countries, four continents
• 358 effect sizes
• Quantitative synthesis through meta-analysis

Business-analytics value configuration

• Technical resources: data, platforms, and analytical tools
• Human resources: analytical and business expertise
• Management capabilities: coordination and strategic alignment
• Complementary conditions: data quality and organizational culture
• Context: external pressure and other firm conditions

Together, these resources and capabilities contribute to firm performance.

This is a variance perspective because the study explains variation in firm performance through variation in organizational resources, capabilities, and contextual conditions.

The paper complements Müller et al. (2018): rather than focusing on objective BDA assets and productivity in one panel dataset, it synthesizes many firm-level studies and asks which BA-related factors matter most across the literature.

Meta-analysis results

Independent variable	n	N	Estimated effect	95% CI	Heterogeneity
BA technical resources	91	39,246	0.452	0.409–0.493	\(Q = 1920.153^{**}\), \(df = 90\), \(I^2 = 95.31\%\)
BA human resources	42	9,219	0.494	0.444–0.541	\(Q = 379.967^{**}\), \(df = 41\), \(I^2 = 89.21\%\)
BA management capabilities	69	16,074	0.477	0.433–0.519	\(Q = 861.169^{**}\), \(df = 68\), \(I^2 = 92.10\%\)
Data quality	14	3,088	0.434	0.328–0.528	\(Q = 142.257^{**}\), \(df = 13\), \(I^2 = 90.86\%\)
Organizational culture	18	3,698	0.472	0.391–0.546	\(Q = 153.939^{**}\), \(df = 17\), \(I^2 = 88.96\%\)
External pressure	17	3,644	0.165	0.036–0.288	\(Q = 248.937^{**}\), \(df = 16\), \(I^2 = 93.57\%\)

Key findings and implications

• BA is positively associated with firm performance across studies.
• The strongest effects are linked to human resources, management capabilities, and organizational culture.
• Technical resources and data quality matter, but they do not explain value creation alone.
• External pressure shows a smaller effect.
• High heterogeneity indicates that BA value creation remains context-dependent.

IV = independent variable; n = number of studies; N = sample size; k = number of correlations; Est. eff. = estimated summary effect size; CI = 95% confidence interval; PI = 80% prediction interval; significance: ** p < .01, * p < .05, not significant (n.s.) for p > .05; \(\tau^2\) = between-study variance; I² = proportion of variance attributed to between-study variance; ^† indicates the number of missing studies necessary for each identified study to nullify the effect, calculated as fail-safe N divided by n; DF = degree of freedom.

Organizational perspectives

A process perspective

Kunz et al. (2025): Process-level value creation from business analytics

This theoretical literature review examines how business analytics creates value in organizational processes and how machine learning changes these value-creation paths.

Research focus

• Process-level value creation
• Business analytics systems in use
• Organizational decision-making
• Knowledge creation and use
• Effects of ML-based analytics

Method and evidence base

• Theoretical literature review
• Emergent theory building
• 176 articles synthesized
• Information Value Chain as sensitizing lens
• Six value-creation paths identified

• RQ 1: How does business analytics create value for organizations?
• RQ 2: How does machine learning change value creation from business analytics?

This slide shifts from variance-based explanations of performance differences to a process perspective: analytics creates value through sequences linking data, information, knowledge, decisions, actions, and process-level outcomes.

Kunz, Spohrer, and Heinzl identify six process-level value creation paths. They also argue that ML changes BA value creation through autonomy, dynamism, and opacity, which strengthens decision automation and feedback loops while complicating knowledge creation and transfer.

Value creation paths

• Ia. Canonical information processing path The emergency department dashboard shows waiting times, patient acuity, bed availability, and historical bottlenecks. The shift lead interprets these data, updates their understanding of the situation, decides to reassign nurses, and acts accordingly.

• Ib. Decision augmentation path A prediction model flags that a patient is likely to deteriorate within two hours. The physician does not fully understand the model’s reasoning but uses the risk score as an additional input alongside clinical judgment.

• Ic. Decision automation path When predicted waiting-room congestion exceeds a threshold, the system automatically sends low-acuity patients to a partner urgent-care clinic and updates appointment slots without human approval.

• IIa. Knowledge accumulation path Over months, the hospital learns that congestion is not mainly caused by emergency arrivals but by delayed discharges from specific wards after 3 p.m. This accumulated insight leads to redesigning discharge routines.

• IIb. Knowledge transfer path Insights from the emergency department are shared with radiology, inpatient wards, and hospital management. Radiology adjusts staffing because ED bottlenecks often follow imaging delays; wards change discharge planning.

• IIc. Feedback path Clinicians notice that the deterioration-risk model overpredicts risk for elderly patients with certain chronic conditions. They report this, the data science team revises features and retrains the model, and future alerts become more relevant.

How machine learning changes value creation

• Autonomy leads to a stronger reliance on analytical information for decision-making and action
• Dynamism involves knowledge and data-driven feedback to improve the ground truth or the analytical system
• Opacity refers to the inscrutability of machine learning models, which restricts human understanding and process improvement

Towards a mature analytics portfolio

The TDWI Analytics Maturity Model offers a diagnostic framework for assessing analytics capabilities across organization, resources, data infrastructure, analytics, and governance.

The model classifies analytics maturity into five stages:

• Nascent → Early → Established → Mature → Advanced/Visionary

TDWI’s model supports a structured assessment of where an organization’s analytics portfolio currently stands and where further development may be needed.

TDWI emphasizes that analytics maturity is not only about techniques or software; it includes organizational processes, culture, resources, data management, and governance. For our purposes, the model is useful because it makes the analytics portfolio assessable across several dimensions and maturity stages. The strategic alignment question is central: are analytics initiatives connected to overall organizational goals, and are they supported by leadership, funding, roles, and governance?

Based on Halper (2020)

People perspectives

A discourse perspective on organizational deployment

EU AI Act: High-risk AI systems

The AI Act establishes specific requirements for AI systems classified as high-risk:

“High-risk AI systems shall comply with the requirements laid down in this Section.”

— Regulation (EU) 2024/1689, Art. 8(1)

High-risk systems must address requirements concerning:

• Risk management — Art. 9
• Data and data governance — Art. 10
• Technical documentation — Art. 11
• Record-keeping and logging — Art. 12
• Transparency and information to deployers — Art. 13
• Human oversight — Art. 14
• Accuracy, robustness, and cybersecurity — Art. 15

High-risk does not mean “prohibited” — it means subject to a comprehensive regulatory regime.

Source: Regulation (EU) 2024/1689 — EU Artificial Intelligence Act, Chapter III, Section 2, Arts. 8–15

EU AI Act: Which systems are high-risk?

Article 6 identifies two main routes to classification as high-risk:

1. AI used as a safety component of certain regulated products
   (Art. 6(1); Annex I)

2. AI used for specified consequential purposes
   (Art. 6(2); Annex III)

Examples of relevant areas:

Area	Examples of high-risk use
Finance	Evaluating a person’s creditworthiness or establishing a credit score — e.g., supporting a loan application decision
Insurance	Risk assessment and pricing for life and health insurance
Recruitment	Filtering applications, evaluating candidates, targeted job advertising
Worker management	Decisions about promotion or termination; task allocation; monitoring and evaluating employee performance
Critical infrastructure	AI used as a safety component in critical digital infrastructure or energy, water, and transport infrastructure

Source: Regulation (EU) 2024/1689 — Art. 6 and Annex III

GDPR: Limits on automated individual decisions

Article 22 — Automated individual decision-making, including profiling

A data subject has a right not to be subject to a decision “based solely on automated processing” when that decision produces legal effects or otherwise similarly significantly affects the person.

Article 22 does not simply prohibit all automation.

Exceptions include decisions that are:

• necessary for entering into or performing a contract,
• authorized by Union or Member State law with appropriate safeguards, or
• based on explicit consent.

Where the relevant exceptions apply, safeguards include at least:

• human intervention
• the opportunity to express one’s point of view
• the opportunity to contest the decision

Article 22 is narrower than a general requirement to make every algorithm explainable. Its core trigger is a decision that is:

1. based solely on automated processing, and
2. has legal or similarly significant effects.

This is a useful distinction for the loan case:

AI Act → regulates the high-risk system and the organizational arrangements around its deployment.

GDPR → protects the individual where personal-data processing and automated decision-making affect them.

Together they motivate the subsequent discussion of human oversight, contestability, and explainability.

Source: Regulation (EU) 2016/679 (GDPR), Art. 22(1)–(4); Recital 71. See also Art. 15(1)(h) on information concerning automated decision-making.

Responsible deployment: Analytics and AI risk management

Advanced analytics increasingly becomes part of AI systems when it informs, recommends, or automates consequential decisions.

NIST AI Risk Management Framework

• Govern: roles, policies, accountability
• Map: context, stakeholders, intended use
• Measure: performance, bias, robustness, risks
• Manage: mitigation, monitoring, incident response

Deployment questions

• Who is affected?
• What decision is supported or automated?
• What can go wrong?
• Who is accountable?
• How do we know when to stop or intervene?

The GDPR (Art. 15 and 22) and the AI Act (Art. 27 and 86) require organizations to support responsible use of consequential automated decisions: affected persons need meaningful information, human intervention and contestability, while deployers of high-risk AI systems need transparency, oversight, logging, and fundamental-rights assessments.

This slide connects the maturity-model idea to responsible analytics. Analytics deployment is not only about technical maturity or portfolio maturity. It also requires risk-management maturity. NIST is useful because it gives students a vocabulary that travels across analytics, machine learning, and AI governance.

• GDPR Art. 15 and Art. 22: automated decision-making Data subjects can receive meaningful information about the logic involved in relevant automated decision-making, including significance and envisaged consequences. In specified cases, safeguards include human intervention, expressing one’s view, and contesting the decision.

• AI Act Art. 13: transparency and instructions for use High-risk AI systems must provide information that helps deployers interpret outputs and use the system appropriately, including intended purpose, performance limits, risks, human oversight measures, and logging mechanisms.

• AI Act Art. 27 and Art. 86: fundamental-rights assessment and explanation For certain high-risk AI uses, deployers must assess fundamental-rights impacts before deployment. Affected persons may also have a right to clear and meaningful explanations of the role of the AI system in relevant individual decisions.

Also: GDPR: processing of personal data: Personal data may only be processed with a lawful basis, such as contract necessity, legal obligation, legitimate interest, or valid consent. In employment settings, consent is often problematic because of the power imbalance. GDPR also requires purpose limitation, data minimization, transparency, retention limits, and accountability.

The black box of modern analytics

Modern analytics systems transform data into predictions without exposing a decision logic that people can easily inspect.

This raises practical questions:

• Why this prediction?
• Why not another prediction?

• When does the system fail?
• Who can contest or override it?

Start by slowing down the assumption that “more predictive performance is always better.” The black-box issue is not simply that the model is mathematically complex; it is that the model may influence decisions about work, credit, health, or other consequential domains, while users and affected persons cannot easily assess whether the output is appropriate.

Use the motor/car prompt here: we do not need to understand combustion physics to drive a car, but we do need indicators, warnings, maintenance rules, legal accountability, and emergency controls. Likewise, XAI supports responsible use, not total technical understanding by every user.

The IUI XAI lecture presents the black-box pipeline, explanation interface, local/global interpretability, intrinsic/post-hoc interpretability, and the notional interpretability–accuracy trade-off. Gunning’s XAI work similarly frames the problem as enabling users to understand, appropriately trust, and manage AI systems.

Sources: https://iui-lecture.org/files/IUI-2021-01-21-slides/IUI_Explainable_AI.pdf ; Gunning (DARPA XAI); Gunning et al. (2019).

The black-box problem “[…] stems from the mismatch between mathematical optimization in high-dimensionality characteristic of machine learning and the demands of human-scale reasoning and styles of semantic interpretation.” Burrell (2016)

The need for interpretability

Interpretability is needed for different purposes

Context	Purpose
Trust	know when to rely on the output
Engineering	debug errors, bias, and brittle behavior
Governance	assign responsibility and define oversight
Recourse	understand, challenge, or improve an outcome
Regulation	document compliance and enable auditability

Emphasize that XAI is a socio-technical control. A technically elegant explanation may still fail if the recipient cannot understand it, cannot act on it, or has no authority to intervene.

Regulatory anchor: GDPR Article 15 includes access to “meaningful information about the logic involved” in automated decision-making in relevant cases; GDPR Article 22 limits solely automated decisions with legal or similarly significant effects and requires safeguards such as human intervention, expression of one’s point of view, and contesting the decision in specified cases. The EU AI Act requires transparency for high-risk AI systems so deployers can interpret outputs and use the system appropriately, includes human oversight requirements, and Article 86 provides a right to clear and meaningful explanations for affected persons in specified high-risk AI cases.

Examples: Reuters reported that Amazon abandoned an experimental internal recruiting tool after it showed bias against women. In credit, adverse-action requirements illustrate the practical need for specific reasons rather than generic model outputs. In health care, Obermeyer et al. showed that an algorithm using cost as a proxy for health need produced racial bias because costs reflected unequal access and spending, not only medical need.

Possible prompt: “Which stakeholder needs which explanation in each example: the applicant/patient, frontline professional, manager, auditor, or engineer?”

Based on Chamola et al. (2023)

Explainability methods

Explainability methods aim to make AI-based systems, their reasoning processes, or their outputs understandable, either by designing models that are explainable from the outset or by generating explanations for already-trained, potentially opaque models.

This slide should provide a map of the main XAI design space, not a technical deep dive. Keep the distinction between explanation families and explanation goals clear.

Feature importance can be global or local. It is intuitive for students, but it can be misleading if interpreted causally. Model-specific examples include coefficients in regression or split criteria in trees; post-hoc examples include LIME/SHAP-style methods that approximate local model behavior. Counterfactuals are attractive because they are action-oriented: they tell an affected person what would have to change for a different outcome. Interactive tools are useful for learning and governance because they allow users to probe thresholds, edge cases, and sensitivity.

In class, ask students which explanation is useful for which stakeholder: an engineer may want debugging information; an applicant may want reasons and recourse; a manager may need limits and oversight rules; an auditor may need logs and evidence.

Sources: Ribeiro, Singh & Guestrin (2016) on LIME; Wachter, Mittelstadt & Russell (2017/2018) on counterfactual explanations; IUI XAI lecture on local/global and intrinsic/post-hoc interpretability.

Based on Speith (2022)

Explainability methods: Examples

Feature importance

Mostly static analysis of the trained model and data; often post-hoc, though some models provide importance intrinsically. Usually provides a global view: which variables matter most across the model or dataset.

Counterfactual explanations

Post-hoc explanation based on minimal feature changes that would alter the prediction.

Interactive explanation

Post-hoc, user-driven exploration combining static explanations with what-if changes to input features.

Human resources for analytics

BCG AI Radar 2025 survey

• 1,803 C-level executives
• Across markets, industries, roles, and company sizes

Main finding: AI is expected to reshape work more than eliminate workers.

• Only 7% expect fewer FTEs due to automation
• 17% expect workforce restructuring with new roles
• 68% expect more productivity and upskilling of the existing workforce
• 8% expect more FTEs and new skills

Analytics and AI deployment is a people-and-process transformation: executives expect value to come from humans and AI working side by side, with human oversight remaining central.

The slide should counter the simplistic “AI replaces people” narrative. The BCG evidence suggests that the central challenge is not only model performance or infrastructure, but organizational change: workflow redesign, incentives, culture, hiring AI talent, and upskilling the existing workforce.

Relevant survey evidence: BCG reports the 10-20-70 principle, with 70% of value creation attributed to people and processes. It also reports that executives see talent and AI as complementary: 64% describe AI and humans as working side by side, 22% see AI taking the lead with humans retaining oversight, and 14% prioritize human talent while using AI only when necessary. On workforce change, only 7% expect fewer FTEs due to AI automation, while 68% expect productivity and upskilling of the existing workforce, 17% expect restructuring with new roles, and 8% expect more FTEs.

Based on Apotheker et al. (2025)

Possible human futures in analytics and AI

Use the upper part to emphasize that analytics adoption depends on human capabilities and organizational acceptance. Skills include technical competence, but also communication, critical thinking, and a mindset of responsible experimentation.

The “adoption tensions” label is deliberately broad. It covers resistance, professional identity, role changes, autonomy concerns, and gaming behavior. This avoids framing resistance as irrational: people may resist because analytics changes status, accountability, incentives, and what counts as expertise.

After the fragment, shift to the future perspective. The point is that “human judgment” is not one stable thing. In some contexts, humans may be removed from the operational decision process. In others, humans become more powerful because analytics extends their capabilities. In still others, the main challenge is uncertainty: humans cannot easily know when AI is strong or weak because performance follows a jagged frontier.

The core takeaway for students: the question is not simply whether humans should remain in the loop. The harder question is which human role is appropriate for which task, under which conditions, and with which safeguards.

Summary

• Analytics can improve organizational performance, but its value depends on context, complementary resources, and how analytics is embedded in organizational processes.

• Value creation requires an analytics portfolio that combines data infrastructure, technical and human resources, management capabilities, and governance aligned with strategy.

• Responsible deployment means asking not only “Can we optimize this?” but also “Should we?” — defining ethical, legal, and organizational boundaries before systems go live.

• As analytics becomes more autonomous, dynamic, and opaque, organizations need interpretability, risk management, and oversight mechanisms that allow people to understand, contest, and intervene.

• The human role is changing rather than disappearing: analytics and AI create value when people have the skills, mindset, and organizational conditions to work effectively with them.

Feedback

Please take 1–2 minutes to reflect on today's session.

Complete the short feedback survey

References

Apotheker, Jessica, Sylvain Duranton, Vladimir Lukic, et al. 2025. “From Potential to Profit: Closing the AI Impact Gap: BCG AI Radar.” Boston Consulting Group, January 15. https://www.bcg.com/publications/2025/closing-the-ai-impact-gap.

Burrell, Jenna. 2016. “How the Machine ‘Thinks’: Understanding Opacity in Machine Learning Algorithms.” Big Data & Society 3 (1): 2053951715622512. https://doi.org/10.1177/205395171562251.

Chamola, Vinay, Vikas Hassija, A Razia Sulthana, Debshishu Ghosh, Divyansh Dhingra, and Biplab Sikdar. 2023. “A Review of Trustworthy and Explainable Artificial Intelligence (XAI).” IEEE Access 11: 78994–9015. https://doi.org/10.1109/ACCESS.2023.3294569.

Halper, Fern. 2020. TDWI Analytics Maturity Model Assessment Guide. TDWI Research. https://go.tdwi.org/rs/626-EMC-557/images/TDWI_Analytics-Maturity-Model-Assessment-Guide_2020.pdf.

Kunz, Pascal C, Kai Spohrer, and Armin Heinzl. 2025. “Process-Level Value Creation from Business Analytics: A Theoretical Literature Review of Value Creation Paths and Changes Induced by Machine Learning.” The Journal of Strategic Information Systems 34 (2): 101902. https://doi.org/10.1016/J.JSIS.2025.101902.

Müller, Oliver, Maria Fay, and Jan Vom Brocke. 2018. “The Effect of Big Data and Analytics on Firm Performance: An Econometric Analysis Considering Industry Characteristics.” Journal of Management Information Systems 35 (2): 488–509. https://doi.org/10.1080/07421222.2018.1451955.

Oesterreich, Thuy Duong, Eduard Anton, and Frank Teuteberg. 2022. “What Translates Big Data into Business Value? A Meta-Analysis of the Impacts of Business Analytics on Firm Performance.” Information & Management 59 (6): 103685. https://doi.org/10.1016/J.IM.2022.103685.

Speith, Timo. 2022. “A Review of Taxonomies of Explainable Artificial Intelligence (XAI) Methods.” Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, 2239–50. https://doi.org/10.1145/3531146.3534639.